3DDPS: A traffic matrix estimation method based on three-dimensional diffusion posterior sampling

Traffic matrix (TM) estimation is an essential but high-cost task for network management. A rational way is to estimate the TMs from the low-cost link load measurements by solving a group of linear equations. However, one open challenge is these linear equations are severely ill-posed in most cases. Fortunately, the emerging deep generative models offer new advanced ways to well address the ill-posed problem. In this paper, we leverage the powerful ability of diffusion models to propose a novel TM estimation framework (named 3DDPS-TME). Different from existing generative-based TM estimation methods, our new method reconstructs the raw TM data into 3D-tensor samples and modifies the typical diffusion framework to 3D-UNet to learn the spatio-temporal correlations of TMs. Furthermore, we adopt diffusion posterior sampling (DPS) for conditional sampling to produce an unbiased TM through a single sampling process. Through extensive experiments and comprehensive comparisons with four state-of-the-art baselines, the experimental results demonstrate that our method exhibits a significant superiority in both estimation accuracy and time consumption. Particularly, using only 0.03%$\sim$10.43% computational cost of the baseline methods, our method makes an improvement of 27%$\sim$68% in terms of estimation accuracy. The codes of the experiments with the proposed methods are available at https://github.com/depositoryL/3DDPS-TME.git.